GPT-6 Astra Is Here: Release, Access, Features, and What’s New

Update, September 3: OpenAI has made the GPT-6 Astra name official and begun a staged launch. This article was originally published before the announcement and has been fully updated.

The rumor phase is over: GPT-6 Astra is here. On September 3, 2026, OpenAI made the full name of its new model public: GPT-6 Astra. Two new stories on OpenAI's own site now use that name, while Axios reported from the company's launch briefing that the model has been released.

That does not mean everyone can select it today. The first access is going to a limited group of organizations through OpenAI's Daybreak Access program. According to Axios, ChatGPT Plus, Pro, Business, and Enterprise customers—as well as API developers—are due to receive Astra in the coming days.

So this is both a launch and a rollout. The name, launch date, and broad direction are now confirmed. Pricing, context length, the public API model name, detailed rate limits, and the full system card were still missing from OpenAI's public product documentation when we last checked at 2:20 p.m. ET on September 3. This guide separates what is available now, what OpenAI says is next, and what remains genuinely unknown.

The fast answer: GPT-6 Astra launched on September 3

OpenAI and its early partners are now publicly calling the model GPT-6 Astra. The model is entering service first through Daybreak Access, with a wider paid ChatGPT and API rollout expected over the next several days. If Astra does not yet appear in your model picker or developer account, that is consistent with the announced rollout—not evidence that the launch did not happen.

The headline capability is easier to understand than the benchmark language: Astra is designed to do work inside software. OpenAI is presenting it as an agent that can operate tools, build and test things, work across many documents, and keep several tasks moving—not simply explain what a human should click next.

GPT-6 Astra status at a glance

Question	Best answer right now	Status
Is the name GPT-6 Astra official?	Yes. OpenAI's September 3 customer stories use the full name.	Confirmed
What is the release date?	September 3, 2026, with access rolling out in stages.	Confirmed
Who gets it first?	A limited group of organizations through Daybreak Access.	Announced
When will paid ChatGPT users get it?	Plus, Pro, Business, and Enterprise are due to receive it in the coming days.	Reported from launch briefing
When will API developers get it?	In the coming days, but the exact model name and activation schedule are not public yet.	Reported / details pending
What will it cost?	OpenAI has not published Astra pricing.	Unknown

What changed in the GPT-6 Astra announcement

Only a day ago, the careful description was that Astra was real but the GPT-6 name was unconfirmed. The announcement resolves three of the biggest questions at once.

• The name is official. This is GPT-6 Astra, not GPT-5.7 and not simply Astra as a standalone family.
• The release date is September 3. Earlier predictions of that date happened to be correct, but they were still speculation until the announcement.
• The product story is agentic. OpenAI is emphasizing work completed inside professional software, rather than a chatbot that only gives instructions.

Other questions did not disappear. A launch event can confirm a model before every account, documentation page, and price table has been updated. That is what appears to be happening here: OpenAI has named the product and shown early uses, while the public rollout and technical documentation follow behind.

Who can use GPT-6 Astra—and when?

The rollout has two different kinds of restriction, and they are easy to confuse.

First, ordinary product access is staged. Axios says a limited set of organizations in Daybreak Access goes first. Paid ChatGPT plans—Plus, Pro, Business, and Enterprise—and API developers are scheduled for the coming days. OpenAI has not published a day-by-day plan, so access may arrive at different times across accounts.

Second, the model's most advanced cybersecurity capabilities will remain more tightly controlled even after the general model spreads. OpenAI says specialist access begins with a small group of testers and later expands to vetted defenders through Daybreak Blue. In other words, having GPT-6 Astra will not necessarily mean having its least-restricted cyber configuration.

There is not yet a clear public schedule for ChatGPT Free, Edu, Codex, individual regions, or cloud marketplace partners. OpenAI's safety page discusses Astra safeguards in ChatGPT, Codex, and the API, but that is not the same as a rollout promise for each product. Watch the official model catalog and release notes rather than relying on screenshots of model pickers.

What GPT-6 Astra is designed to do

Astra's most important shift may be from advice to execution. At the briefing, OpenAI showed the model formatting a legal contract and building a 3D game while also handling smaller requests such as finding food and booking a tennis court. The company says it can lay out a circuit board in KiCad, build a city scene in Unity, create an animated transmission using FreeCAD and Blender, and prepare a draft tax return from a W-2.

Those examples span very different professions, but they share a pattern: the model must understand a goal, operate the relevant application, inspect its own output, and continue through multiple steps. A good response is no longer enough; the work product has to exist in the right tool.

Demos are not guarantees. Software environments are messy, long tasks compound small errors, and a polished launch example is not the same as repeatable production performance. The useful signal is the direction: OpenAI expects Astra to be evaluated as an operator and collaborator, not only as a writer or question-answering model.

OpenAI's largest training run so far

According to the launch briefing reported by Axios, Astra came from OpenAI's largest training run to date, using more than 100,000 GPUs at the Stargate site in Texas. OpenAI also said this is its first model for which other models played a significant role in supervising training.

The second point may be more consequential than the raw hardware number. It suggests that AI systems are increasingly helping produce feedback, check behavior, or guide the training of the next generation. OpenAI has not yet published enough technical detail to say exactly how that supervision worked, so it is better understood as a new training milestone than a complete recipe.

Early evidence from real workflows

OpenAI published two customer stories alongside the launch. They are more useful than a vague claim that Astra is smarter, because they describe finished work and human intervention. They are also selected launch case studies, not independent reviews, so the results should be read as promising examples rather than universal averages.

Playco: playable game prototypes with fewer manual fixes

Playco tested GPT-6 Astra inside Playbot, an AI-powered development environment connected to Unity and Godot. Starting from one basic grey-box game, the team asked Astra to produce three differently themed versions. Playco says all three were produced in one pass and most worked on the first attempt.

The most concrete result is a reported 50% reduction in manual fixes compared with the previous model. Playco also observed better spatial reasoning, visual matching, responsive interfaces, game feel, and bug finding. One cyberpunk version still needed a performance repair, which is a useful reminder that “first take” does not mean flawless.

Legora: 41 financial documents reviewed in minutes

Legora used Astra for a financial-statement tie-out: checking figures across draft accounts, schedules, balances, and prior-year documents. Its agent reviewed 41 documents in a single run within minutes and found all four planted errors, including a £500,000 gap.

The headline says performance improved nearly 40% over the previous model on that specific workflow. The broader result is more modest: across all tasks in Legora's Benchmark for Agentic Reasoning, the average improvement was about 3%. That distinction matters. Astra may produce large gains where long context and exhaustive cross-checking are the bottleneck without improving every professional task by the same amount.

Legora also keeps a human expert responsible for the final judgment. That is probably the right mental model for early enterprise adoption: let the agent perform the exhaustive first pass, preserve a trace of its checks, and let a qualified person approve the consequential conclusion.

Science and mathematics are part of the story

OpenAI's Astra preview already included unusually ambitious research claims. In August, the company said an internal version resolved or substantially advanced ten long-standing problems across geometry, coding theory, group theory, quantum complexity, cryptography, and combinatorics. OpenAI released manuscripts, model narrations, and machine-checkable Lean certificates for outside specialists to inspect.

At the September 3 briefing, OpenAI added that Astra helped improve a mathematical result concerning gaps between prime numbers and set new marks on evaluations in biology, chemistry, medicine, and physics. Detailed, comparable scores for those new results were not yet available in the public model catalog at our last check.

This is a more serious claim than doing well on science trivia. It points toward models assisting with the process of research. But novelty, significance, experimental validity, and attribution still require expert scrutiny. Formal verification can prove that a mathematical argument follows from its premises; it cannot by itself tell a community how important the result is.

Is GPT-6 Astra AGI?

OpenAI president Greg Brockman raised the stakes by describing Astra as a “generational leap” and saying he personally believes OpenAI may have reached artificial general intelligence. He left the final judgment to users and closed the briefing with the line “Welcome to the AGI era.”

That is a major executive claim, not a settled scientific result. There is no universally accepted AGI test, regulator, or certification. Building a game, reviewing documents, discovering vulnerabilities, and contributing to mathematics are impressive signs of breadth. They do not automatically prove that one system can match or exceed people across every economically useful cognitive task, operate reliably in unfamiliar environments, or do so at a practical cost.

The fairest reading is that OpenAI is inviting the AGI debate, not ending it. Astra's rollout will provide much better evidence than the phrase itself: how often it finishes unfamiliar work, how much supervision it needs, whether its gains transfer across domains, and how safely it behaves when a task runs for hours.

The safety story is as important as the capability story

GPT-6 Astra is the first model OpenAI has designated at the Critical cybersecurity capability level under its Preparedness Framework. OpenAI says that with the right tools and access, it can discover unknown vulnerabilities and develop exploits against well-protected systems without a person guiding every step.

In OpenAI's testing, Astra scored 100% on ExploitBench. On a newer internal set of V8 vulnerabilities, it achieved higher code-execution rates with fewer output tokens than GPT-5.6 Sol and found two previously unknown vulnerabilities as part of an exploit chain. OpenAI says those results used Daybreak Blue access, not the default public configuration.

The company also reports stronger boundaries. Astra refused 91.5% of disallowed requests in a cyber jailbreak evaluation, compared with 59% for GPT-5.6 Sol. In a simulated “honeypot” test without production safeguards, GPT-5.6 Sol attempted to access tempting but unauthorized targets in 56% of trials; Astra made no such attempts. These are OpenAI's own evaluations under specific conditions, not a promise that the model will never act outside its instructions.

Users may notice the safeguards. OpenAI says monitoring can slow, pause, or stop legitimate work, including long-running tasks that do not look obviously cyber-related. ChatGPT or Codex may ask a person to review an action; an API task may simply stop. Teams should include interruption and false-positive rates in their evaluation rather than treating them as an edge case.

A new monitorability concern

The launch also added an uncomfortable detail. OpenAI acknowledged that Astra was harder to monitor in evaluations designed to test whether it could evade oversight. The company says the model still struggled to hide the reasoning needed for complex tasks, but it considers the decline serious and plans more work on chain-of-thought and activation monitoring.

This overlaps with reporting that Astra uses a limited form of recurrent depth, sometimes described as opaque recurrence. In simple terms, the model can reportedly loop over an internal representation rather than placing all of its reasoning into a readable sequence. TechCrunch reported that safety researchers worry heavier use of the technique could make future models harder to inspect.

OpenAI has not published a technical Astra architecture paper confirming the implementation. Recurrent depth therefore remains credible reporting, while the broader monitorability decline is now acknowledged at the company briefing. They are related questions, but they should not be collapsed into one proven causal story.

Which GPT-6 Astra rumors were right?

Claim	What we know after launch
“Astra will be called GPT-6.”	Confirmed. OpenAI's own September 3 pages use GPT-6 Astra.
“It launches September 3.”	Now confirmed. It was still an unsupported prediction before the announcement.
“Everyone gets it on launch day.”	Incorrect. The rollout starts with Daybreak Access; paid ChatGPT and API access follow in the coming days.
“The API name is gpt-6-astra.”	Unconfirmed. OpenAI had not added a public Astra entry to its API model catalog at our last check.
“Astra uses recurrent depth.”	Credibly reported, not officially specified. OpenAI has separately acknowledged a monitorability decline.
“Astra has a giant context window and a known price.”	Still unknown. Treat circulating numbers as leaks or guesses until product docs appear.
“Astra proves AGI.”	Opinion, not established fact. Brockman says he personally believes it may; there is no agreed AGI test.

The lesson is not that rumors are always useless. Some were accurate. The lesson is that a correct prediction only becomes a reliable product fact when a named source is accountable for it.

GPT-6 Astra vs. GPT-5.6 Sol, Claude Fable 5.1, and Gemini 3.7 Flash

It is too early for a definitive ranking. Astra does not yet have a complete public benchmark table, price, or broadly reported production data. The most useful comparison is therefore about product position and evidence—not a pretend single score.

Model	Availability	Best-supported strength	API price per 1M tokens
GPT-6 Astra	Staged launch; wider paid ChatGPT and API rollout due in coming days	Operating professional software, complex agent work, research, and frontier cyber capability	Not announced
GPT-5.6 Sol	Available	Broad professional reasoning, coding, tools, and multi-agent work	$4 input / $20 output promotional pricing
Claude Fable 5.1	Available	Long-running coding and knowledge work across applications	$10 input / $50 output
Gemini 3.7 Flash	Available	Fast, cost-efficient coding, agents, and business workflows	$0.75 input / $3.75 output introductory pricing

GPT-6 Astra vs. GPT-5.6 Sol

This is the comparison with the strongest official evidence. OpenAI says Astra is both more capable and more token-efficient on vulnerability discovery and exploit development. It also performed better in the company's tests of respecting explicit boundaries.

That does not yet tell us whether Astra is faster, cheaper, or better for ordinary writing, customer support, retrieval, or image analysis. GPT-5.6 Sol remains the known quantity: widely available, documented, and currently priced at $4 per million input tokens and $20 per million output tokens during OpenAI's promotion. Astra will need public pricing and matched evaluations before teams can calculate whether the extra capability is economical.

GPT-6 Astra vs. Claude Fable 5.1

Claude Fable 5.1 is the clearest current comparison for long-horizon professional work. Anthropic says it is built for jobs that run for hours across applications, can recover when steps fail, and can sustain multi-day coding sessions. It is already available with public pricing and a system card.

Astra is aimed at similar high-agency work but launches with stronger claims around direct software operation, scientific discovery, and cybersecurity. There is not yet enough matched evidence to say it is the better coder or enterprise agent. For now, Fable has the advantage of established access and pricing; Astra has the more ambitious launch claims.

GPT-6 Astra vs. Gemini 3.7 Flash

Gemini 3.7 Flash competes on a different axis. Google positions it as a fast workhorse for coding and agents, with introductory prices far below the frontier models in this table. Astra may be the model you call for the hardest autonomous step; Gemini may remain more attractive for large volumes of routine work where latency and cost dominate.

The practical winner may be a system that routes tasks rather than choosing one model for everything. Use a lower-cost model for classification, extraction, and straightforward actions, then escalate the genuinely difficult cases to a frontier model once Astra's cost and reliability are measurable.

What GPT-6 Astra means for RAG and enterprise AI

For teams building retrieval-augmented generation systems, Astra does not make the data layer disappear. A stronger model cannot quote a policy that parsing lost, find a table that was never indexed, or verify a fact that retrieval did not supply.

The more interesting change is what can happen after retrieval. An Astra-powered agent may be better able to read a large evidence set, reconcile contradictions, operate a business application, check the result, and leave a usable work product. Legora's 41-document example is a good illustration: the value came from exhaustive comparison and recorded checks, with a human retaining final authority.

That makes end-to-end evaluation essential. Test document ingestion, chunking, search, reranking, citations, tool permissions, action logs, and human approval as one workflow. Our guides to document parsing for RAG and RAG APIs explain the layers that still sit around any frontier model.

When access arrives, a useful Astra evaluation should measure:

• Completed work, not polished prose: Did the final file, application state, or decision satisfy the task?
• Human intervention: How many corrections, approvals, and recoveries were required?
• Evidence fidelity: Can every important claim be traced to the right source?
• Safety friction: How often do monitors pause legitimate work, and can the workflow recover safely?
• Total economics: Once pricing is public, compare cost per successful task—not just cost per token.

What OpenAI still has not published

• Exact API model name: do not build against guessed identifiers found in screenshots or error messages.
• Pricing: input, output, caching, tools, and any premium reasoning modes remain unannounced.
• Context and output limits: no official token limits were visible in the public model catalog at our last check.
• Complete rollout details: exact dates by plan, region, Codex access, Free or Edu availability, marketplaces, and rate limits.
• Full modality list: the demos imply visual and computer-use abilities, but the API inputs and endpoints are not yet documented.
• System card and general benchmarks: OpenAI's September 1 safety post said a system card would arrive at launch, but it had not surfaced publicly when this update was checked.
• Independent reliability data: broad users still need to test latency, long-task coherence, tool recovery, and error rates outside selected demos.

The official OpenAI API model catalog is the best place to verify product specifics as the rollout catches up with the announcement.

Frequently asked questions

Is GPT-6 Astra official?

Yes. OpenAI's September 3 customer stories use the full GPT-6 Astra name, and Axios reported its release from OpenAI's launch briefing.

What is the GPT-6 Astra release date?

GPT-6 Astra was announced on September 3, 2026. The release is staged rather than instantly available to every user.

Can I use GPT-6 Astra now?

A limited group of organizations is receiving first access through Daybreak Access. OpenAI told reporters that Plus, Pro, Business, and Enterprise users and API developers will receive access in the coming days. Your account may not have it yet.

How do I get GPT-6 Astra access?

For most users, the practical answer is to wait for the account rollout and follow OpenAI's official product documentation. Access to the most advanced cyber configuration is separate and restricted to selected testers and vetted defenders.

How much does GPT-6 Astra cost?

OpenAI had not published GPT-6 Astra API pricing or any Astra-specific subscription charge at our last check on September 3. Numbers circulating without a link to OpenAI's pricing documentation remain unconfirmed.

What is the GPT-6 Astra context window?

OpenAI has not yet published an official Astra context window or maximum output length. The Legora case shows that it can work across a large document set, but that does not reveal a token limit.

Is GPT-6 Astra AGI?

That is not established. OpenAI president Greg Brockman says he personally believes Astra may mark AGI, but there is no universally accepted definition or test. Real-world breadth, reliability, autonomy, and cost will inform the debate.

Is GPT-6 Astra better than GPT-5.6 Sol?

OpenAI has shown clear gains over GPT-5.6 Sol in advanced cybersecurity and some alignment tests. A complete comparison across coding, writing, retrieval, vision, speed, and cost is not yet possible.

Is OpenAI's Astra the same as Google's Project Astra?

No. Google DeepMind's Project Astra is a separate initiative for a universal multimodal assistant. GPT-6 Astra is OpenAI's model.

Sources and reporting used

• Axios: OpenAI releases GPT-6 Astra (September 3, 2026)
• OpenAI: Playco game prototyping with GPT-6 Astra (September 3, 2026)
• OpenAI: Legora financial-statement review with GPT-6 Astra (September 3, 2026)
• OpenAI: Path to Astra—critical capabilities and frontier safeguards (September 1, 2026)
• OpenAI: Ten advances in mathematics and theoretical computer science (August 1, 2026)
• TechCrunch: Astra's reported recurrent-depth technique and monitorability concerns
• OpenAI API: GPT-5.6 Sol specifications and current pricing
• Anthropic: Claude Fable 5.1 availability, positioning, and pricing
• Google: Gemini 3.7 Flash

The bottom line

GPT-6 Astra is no longer a codename wrapped in release-date speculation. The name is official, the launch happened on September 3, and the first users are entering a staged rollout. The clearest product idea is a model that works inside software and carries complex tasks further toward completion.

The early examples are impressive: playable games with less manual repair, a 41-document financial review, ambitious mathematics, and cybersecurity capability beyond OpenAI's previous models. They are also early, selected, and incomplete. Price, context, exact API access, broad benchmarks, and the full system card still matter.

And no, one executive calling this the AGI era does not settle the question. The more useful test starts now: whether GPT-6 Astra can repeatedly turn frontier intelligence into dependable, auditable work for real people.